adapt(musa): MTT S5000 backend adaptation for vLLM 0.24.0

[physics31415926]

Summary

Upgrade compatibility patches for MTT S5000 (MUSA) offline inference under vllm 0.24.0, covering both dense and MoE models.

Environment

Component	Version
Docker image	registry.mthreads.com/mcconline/inference/vllm:v0.20.0-ph1-4.3.5-torch2.9-20260605-mtcc51
vllm	0.24.0
torch / torch_musa	2.9.0
flagtree (triton)	0.6.0+mthreads3.6
FlagGems	mthreads branch
Hardware	MTT S5000 × 4 (TP=4)

Changes

File	Fix	Root Cause
musa/patch.py	patch_accelerator_missing_attrs: unconditionally replace torch.accelerator.empty_cache with torch_musa.empty_cache; add max_memory_allocated and device_index patches	Built-in impl calls _accelerator_isAllocatorInitialized() which asserts on MUSA; max_memory_allocated called by base_loader after model load; device_index used by fla/ops
musa/patch.py	patch_cuda_get_device_properties: fix sentinel flag name	Stale sentinel _musa_get_device_capability_patched caused patch to be skipped on reload
musa/patch.py	patch_cuda_stream_for_musa: aux_stream() returns None on MUSA; force-import all consumer modules and patch their local bindings	SharedExperts.__init__ calls aux_stream() → torch.cuda.Stream() (MUSA dummy, raises RuntimeError). Consumer modules use from vllm.utils.torch_utils import aux_stream (local binding at import time), so patching the module attribute alone is insufficient
musa/patch.py	patch_inductor_triton_for_musa: fix self.binary → self.kernel	TritonCompileResult renamed compiled kernel attribute from .binary to .kernel in torch 2.9; raised AttributeError during KV-cache warmup
musa/patch.py	patch_moe_topk_softmax_for_musa: replace threading/meta-path-finder with direct importlib.import_module; add return topk_weights, topk_ids	Threading approach had a race condition (patch thread not guaranteed to complete before first call). Missing return caused TypeError: cannot unpack non-iterable NoneType object
musa/musa.py	Add FlagGems topk_softmax import stubs for MUSA dispatch	Required for MoE router dispatch registration
musa/register_ops.py	Register topk_softmax dispatch for MUSA backend	Required for MoE router dispatch registration

Test Results

All tests run with MUSA_VISIBLE_DEVICES=4,5,6,7, TP=4.

Qwen3.6-27B (dense)

Prompt	Output	Status
Hello, my name is	John. I am a 25-year-old...	✅
The capital of France is	Paris...	✅
What is 2+2?	2+2 equals 4...	✅

Qwen3.6-35B-A3B (MoE)

Prompt	Output	Status
Hello, my name is	John. I am a 25-year-old male...	✅
The capital of France is	Paris, a city renowned for its iconic landmarks...	✅
What is 2+2?	2+2 equals 4...	✅

Both models initialize and exit cleanly with no errors.